What Public Research Shows
The European Broadcasting Union, working with public service media organizations across 18 countries and 14 languages, reported that almost half of evaluated AI assistant news answers had at least one significant issue, with sourcing and accuracy among the major failure areas.
The Reuters Institute Digital News Report 2025 found that 58% of respondents across 48 markets remained concerned about telling what is true from what is false online. It also found that trusted news brands and official sources remain central places people go to check claims.
Where Drift Shows Up Fast
Drift was most visible in claims that sounded plausible but contained subtle timeline or denominator errors. In those cases, output language remained highly confident while source quality degraded.
The Safe Use Pattern
Teams got the best outcomes when they forced three checks before accepting a claim: source recency, source authority, and contradiction scan. This reduced high-confidence errors substantially and made reviewer handoff faster.
For organizations publishing at speed, the winning pattern was not one model versus another. It was staged verification with explicit confidence penalties when evidence was weak or stale.
Source Trail
Sources checked for this brief: EBU News Integrity in AI Assistants, Reuters Institute Digital News Report 2025, and the NIST AI Risk Management Framework.
Bottom Line
AI assistants can help with first-pass synthesis, but publication-grade verification needs stricter evidence discipline than chat ergonomics naturally provide. Use model output inside an evidence-first pipeline with source trails, recency checks, contradiction scans, and human review for high-stakes claims.